Knowledge-gap reports: how AI support improves your help centre
What should you know about AI support knowledge gaps and how a knowledge-gap report fixes them?
Every question the agent cannot answer becomes a ranked list of help-centre gaps, turning failures into content work. The report clusters unanswered and escalated questions by topic, ranks them by volume and cost, and assigns each to an owner, so the help centre improves weekly instead of when someone remembers.
AI support knowledge gaps are the questions your agent could not answer because the help centre, policy documents or account data did not contain the answer. Left alone they are just failures. Logged, clustered and ranked, they become the most accurate content roadmap a support organisation has ever had, because each item is a real question, asked by real customers, with a count next to it. This article explains how a knowledge-gap report is built, what it should contain, who acts on it, and why it is the mechanism that makes an AI support deployment improve month on month rather than plateau after launch.
Why AI support knowledge gaps are the improvement engine
A customer service agent answers from what it can retrieve. When retrieval returns nothing relevant, a well-built agent says so and escalates rather than improvising. That moment is precisely recorded: the question, the intent, the passages that were retrieved and rejected, and what the human eventually said. Multiply by a month of traffic and you have a ranked list of what the help centre is missing, written in the customers' own words. Human-only support teams never had this, because reps answered from memory and the gap was never logged.
What goes into a knowledge-gap report
| Column | Source | What it tells the content owner |
|---|---|---|
| Question cluster | Semantically grouped unanswered and escalated questions | The topic, phrased as customers phrase it |
| Volume | Count over the reporting period | Priority |
| Gap type | Missing article, outdated article, contradictory sources, data not available | What kind of fix is needed |
| Nearest existing content | The best passage retrieved, with its relevance score | Whether to edit or write new |
| Human answer | What the rep said when it escalated | A draft of the fix, already written |
| Cost | Handling time of the escalations in this cluster | The business case for fixing it this week |
| Owner and status | Assigned team, due date, done | Accountability |
The four kinds of support content gaps
Missing
Nobody wrote it. Common for new features, edge cases and anything the product team assumed was obvious. The fix is a new article, and the rep's escalation reply is usually most of the draft.
Outdated
The article exists and is wrong. Prices changed, a menu moved, a policy was updated in the legal document but not the help centre. The agent retrieves it with high confidence and gives a wrong answer, which is worse than no answer. These are detected from reopens and rep corrections rather than from retrieval failures, so the report needs both sources.
Contradictory
Two articles disagree, or the help centre disagrees with the policy PDF. The agent retrieves both and either picks one or hedges. The fix is a single source of truth with the other retired, and this is the gap type that most often needs a decision from a manager rather than a writer.
Data, not content
The answer is not in any document; it is in a system the agent cannot see. "When will my refund reach my bank?" needs the payment gateway's status, not an article. These gaps go to engineering as integration requests, and separating them from content gaps stops the help-centre team being blamed for an API that does not exist yet.
Help centre optimisation AI: from report to fix
The report is only useful if someone acts on it weekly. The routine we install is a thirty-minute review with the support lead and the content owner: read the top ten clusters, classify each, assign it, and set a due date. Fixed items are re-indexed the same day. The following week's report shows whether the cluster shrank. Over a quarter, the top of the list turns over completely, which is the visible sign the loop works. The agent's evaluation set gets a new question from each fixed cluster, so the fix is protected against regression; the method is described in How to measure RAG quality.
Where the agent drafts for reps rather than answering directly, edits are the second signal. Reps correcting the same fact repeatedly is an outdated-article gap even if retrieval never failed; see agent-assist for the edit-review loop.
Ranking the unanswered questions report
Volume alone over-weights cheap questions. Rank by volume multiplied by the average handling time of the escalations in that cluster, which approximates the cost of the gap. A question asked forty times that takes a rep two minutes matters less than one asked ten times that takes twenty minutes and a call-back. Add a manual boost for anything with a compliance or safety angle. Present the top ten only; a report with two hundred rows is not read.
Keep the full list available for anyone who wants it, but the weekly review works from the top ten. If an item stays in the top ten for three consecutive weeks without an owner, that is an organisational problem rather than a content one, and the support lead should raise it as such.
Writing content the agent can use
Help centres written for humans skimming a page are often poor retrieval sources. Content written with the agent in mind is structured differently: one question per article or section, the answer in the first paragraph, exact product terms rather than marketing names, tables for anything with conditions, a date and an owner, and no "see above". The help centre guidance from Zendesk covers the human side well; the retrieval side is covered in Why basic RAG fails in production. Content that is good for the agent turns out to be good for customers searching the help centre too.
Multilingual gaps
If the agent serves several languages, the report is per language. A question that is well covered in English and unanswered in Hindi is a translation gap, and the report should show the English article that needs translating and the volume in each language. The multilingual support guide explains the two-layer knowledge base this works with.
A worked example
A consumer electronics brand launched a support agent over a help centre that had grown by accretion for years. The first knowledge-gap report showed the largest cluster was warranty claims for a product line launched after the warranty article was last edited; the second was contradictory return windows between the help centre and the terms page; the third was "where is my replacement", which was a data gap because replacements were tracked in a separate system the agent could not see. The content team fixed the first two in a week using the reps' escalation replies as drafts. Engineering added a read-only lookup for replacements in the following sprint. The next month's report had none of the three at the top, and the agent's resolution on warranty questions rose without a single prompt change. After a quarter, the support lead described the report as the first time the help centre had a backlog that came from customers rather than from guesses.
Team and timeline
The knowledge-gap report is built into every customer service agent we deliver and into retrieval and knowledge engineering projects generally; it is not a separate purchase. The engineering is a few days: logging retrieval failures with context, clustering weekly, and a dashboard or scheduled export. The organisational part is the harder one and needs a named content owner and a support lead who will hold the weekly review. Under a Care Plan we run the clustering, re-index fixed content and add evaluation cases; the Standard plan at $2,500 / ₹1,60,000 a month covers this for most teams. Starting prices for the agent build are on the pricing page.
Before you start: a checklist
- Confirm the agent logs unanswered questions with the retrieved passages and the eventual human answer
- Decide the reporting period (weekly is right for the first quarter)
- Name a content owner and a support lead for the review
- Agree the ranking formula: volume times handling time, with a manual boost for compliance topics
- Set the rule for data gaps: they go to engineering, not the content team
- Rewrite the top ten articles in agent-friendly structure as a pilot
- Add one evaluation question per fixed gap
- Decide how fixes are re-indexed and how quickly
Questions clients ask
- Can the agent write the missing articles itself? It can draft from the rep's escalation reply and the retrieved context, and a human should review before publishing. Publishing unreviewed drafts creates the outdated and contradictory gaps you are trying to remove.
- How many gaps will we see in month one? Usually many. The number is not the point; the turnover at the top of the list is.
- Does this replace help-centre analytics? No. Search analytics show what people looked for; the gap report shows what the agent could not answer, which is the sharper signal.
- Will the list ever be empty? No. Products change and customers invent new questions. A healthy report has a short, changing top ten.
Related reading
AI ticket deflection is the wrong metric shows where knowledge gaps closed sits in the monthly report, the in-app copilot case study describes a documentation loop in a B2B product, and the pricing page has the Care Plan details.
Log every question the agent cannot answer, rank the clusters by cost, fix the top ten weekly, and the help centre becomes the one part of support that gets better on its own schedule.
Frequently asked questions
What is a knowledge-gap report?
▾
A ranked list of questions the AI support agent could not answer, clustered by topic with volume, cost, the nearest existing content and the human's eventual reply, assigned to an owner for fixing.
Who should own the knowledge-gap review?
▾
A support lead and a help-centre content owner together, weekly for the first quarter. Data gaps that need an integration go to engineering.
How does fixing a gap improve the agent?
▾
The new or corrected article is re-indexed, the agent retrieves it on the next similar question, and an evaluation case is added so a later change cannot silently break it.